All pull requests

Pull requests

• Move the MOPD example to its own folder and add a README

  #8101·sergiopaniego opened Nov 3, 2026Member·

  ·

• Run CI when a draft PR is marked ready for review

  🔧 maintenance
  #8100·albertvillanova opened Nov 3, 2026Member·

  ·
  1

• Add a README with the results to the Wordle example

  #8099·sergiopaniego opened Nov 3, 2026Member·

  ·
  1

• Update the Sudoku example to start from an SFT model with a progress reward

  #8097·sergiopaniego opened Nov 3, 2026Member·

  ·
  1

• Make the AsyncGRPO rollout loop public

  #8096·AmineDiro opened Nov 3, 2026Member·3/3·

  ·
  3

• Load the Countdown examples in the GOLD tests from Countdown-Task-GOLD

  🔧 maintenance
  #8091·albertvillanova opened Nov 3, 2026Member·

  ·
  1

  1

• Skip the experimental LFM2.5-VL test on transformers<5.0.0

  🔧 maintenance
  #8088·albertvillanova opened Nov 3, 2026Member·

  ·
  1

• Skip experimental Gemma4 tests on transformers<5.5.0

  🔧 maintenance
  #8087·albertvillanova opened Nov 3, 2026Member·

  ·
  1

• Skip experimental Qwen3-VL tests on transformers<4.57.0

  🔧 maintenance
  #8086·albertvillanova opened Nov 3, 2026Member·

  ·
  1

• Rewrite the gpt-oss example: fine-tune on chess puzzles, then back to MXFP4

  #8085·qgallouedec opened Nov 3, 2026Member·

  ·
  1

• Derive the training template pairs in the response parsing test

  🔧 maintenance
  #8084·albertvillanova opened Nov 3, 2026Member·

  ·
  1

• Drop the removed core extra from the openenv install tip

  📚 documentation
  #8083·albertvillanova opened Nov 3, 2026Member·

  ·
  1

• Skip the other adapters' modules_to_save copies in the vLLM weight sync

  #8079·qgallouedec opened Nov 2, 2026Member·

  ·
  1

  9

• Stop refusing padding_free=True with max_length in SFT

  #8077·qgallouedec opened Nov 2, 2026Member·

  ·
  1

  2

• Split the AsyncGRPO rollout loop into generate, reward, advantage and…

  #8073·AmineDiro opened Nov 2, 2026Member·2/3·

  ·
  3

• Keep the pre-filled thinking block out of the Qwen3 assistant-only loss

  #8072·albertvillanova opened Nov 2, 2026Member·

  ·
  1

  1

• Load MXFP4 and FP8 checkpoints dequantized by default

  #8070·qgallouedec opened Nov 2, 2026Member·

  ·
  9

• Use the tiny GLM-5.3-Flash model in the GLM think-prefill test

  🔧 maintenance
  #8069·albertvillanova opened Nov 2, 2026Member·

  ·

• Add a tiny GLM-5.3-Flash generation script

  🔧 maintenance
  #8068·albertvillanova opened Nov 2, 2026Member·

  ·
  1

• Drop the rollout_worker argument from AsyncDistillation

  #8066·AmineDiro opened Nov 2, 2026Member·2/2·

  ·
  3

• Drop the weight_transfer protocol from AsyncGRPO and AsyncDistillation

  #8065·AmineDiro opened Nov 2, 2026Member·1/2·

  ·
  1

  3

• Log AsyncGRPO metrics under the GRPO names

  #8064·AmineDiro opened Nov 2, 2026Member·

  ·

• Asyncgrpo config renames

  #8063·AmineDiro opened Nov 2, 2026Member·1/3·

  ·
  3

  1

• Quantize the tiny DeepSeek-V3 and GPT-OSS models like their references (FP8, MXFP4)

  🔧 maintenance
  #8058·qgallouedec opened Nov 1, 2026Member·

  ·
  12

• Fix GOLD cross-tokenizer alignment and teacher inputs

  #8057·kashif opened Nov 1, 2026Collaborator·4 of 8 tasks·

  ·